v0.4.1

Features

• Add opentelemetry to report tensor ready order (#42)

v0.4.0

⚠ BREAKING CHANGE

• BaguaBucketPy::append_decentralized_synchronous_op now only supports full precision decentralized communication.

Features

• make full precision decentralized op stateless (#36) 98319c9
• add (scatter, gather, scatter_reduce) and all inplace version communication primitives (#37) f931473
• add native low precision decentralized operator (#26) 50295e8
• better debug log including tensor info when executing ops 1bd6e0b

Fixes

• substract overflow error for decentralized op (#39) 30cdb67
• fix scatter and reduce_scatter implementation (#40) ee90376
• make compress/decompress of BaguaTensor method string consistent (#33) ee929df

v0.4.0-rc1

Format Rust code using rustfmt

v0.3.1

Fixes

• always mark bagua padding tensor as ready 63f88d4
• fix alltoall_v parameter types (#27) b541d85

v0.3.0

v0.3.0

⚠ BREAKING CHANGE

• BaguaBucketPy and BaguaTensorPy now require name. BaguaTensorPy is created by passing pytorch tensor directly now.

Features

• Compatible mode for getting pytorch tensor info with Python interpreter 1534d23
• Support creating BaguaTensor by passing torch tensor directly (#19) 4306e94
• Support Reduction op selection (SUM, MAX, etc.) b1bf784
• Add reduce and allgather python interface ff68a61
• Support alltoall_v (vector alltoall) (#14) a6fe110

Fixes

• fix python gil deadlock during getting data ptr 6ba6ace
• fix decompress incorrect pointer and typo in error msg a7e34ba

v0.2.1

Features

• add 5 min timeout for buckets' comm op d83005d

v0.2.0

⚠ BREAKING CHANGE

• set_xxx_ops are renamed to append_xxx_op now.

Features

• initial support for python op (#2) 4899f50

v0.1.3

Fixes

• merge bagua_install_library and setup.py, remove nccl<=2.6 support bda550c

v0.1.2

Fixes

• add version.py placeholder to prevent file not found error c35f979

v0.1.1

Features

• install nccl deps in bagua core and add generated version variable 57b90ed

Fixes

• fix ci pypi versioning 62c7012